1 //===- InductiveRangeCheckElimination.cpp - -------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // The InductiveRangeCheckElimination pass splits a loop's iteration space into
10 // three disjoint ranges.  It does that in a way such that the loop running in
11 // the middle loop provably does not need range checks. As an example, it will
12 // convert
13 //
14 //   len = < known positive >
15 //   for (i = 0; i < n; i++) {
16 //     if (0 <= i && i < len) {
17 //       do_something();
18 //     } else {
19 //       throw_out_of_bounds();
20 //     }
21 //   }
22 //
23 // to
24 //
25 //   len = < known positive >
26 //   limit = smin(n, len)
27 //   // no first segment
28 //   for (i = 0; i < limit; i++) {
29 //     if (0 <= i && i < len) { // this check is fully redundant
30 //       do_something();
31 //     } else {
32 //       throw_out_of_bounds();
33 //     }
34 //   }
35 //   for (i = limit; i < n; i++) {
36 //     if (0 <= i && i < len) {
37 //       do_something();
38 //     } else {
39 //       throw_out_of_bounds();
40 //     }
41 //   }
42 //
43 //===----------------------------------------------------------------------===//
44 
45 #include "llvm/Transforms/Scalar/InductiveRangeCheckElimination.h"
46 #include "llvm/ADT/APInt.h"
47 #include "llvm/ADT/ArrayRef.h"
48 #include "llvm/ADT/None.h"
49 #include "llvm/ADT/Optional.h"
50 #include "llvm/ADT/PriorityWorklist.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallVector.h"
53 #include "llvm/ADT/StringRef.h"
54 #include "llvm/ADT/Twine.h"
55 #include "llvm/Analysis/BlockFrequencyInfo.h"
56 #include "llvm/Analysis/BranchProbabilityInfo.h"
57 #include "llvm/Analysis/LoopAnalysisManager.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/ScalarEvolution.h"
60 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
61 #include "llvm/IR/BasicBlock.h"
62 #include "llvm/IR/CFG.h"
63 #include "llvm/IR/Constants.h"
64 #include "llvm/IR/DerivedTypes.h"
65 #include "llvm/IR/Dominators.h"
66 #include "llvm/IR/Function.h"
67 #include "llvm/IR/IRBuilder.h"
68 #include "llvm/IR/InstrTypes.h"
69 #include "llvm/IR/Instructions.h"
70 #include "llvm/IR/Metadata.h"
71 #include "llvm/IR/Module.h"
72 #include "llvm/IR/PatternMatch.h"
73 #include "llvm/IR/Type.h"
74 #include "llvm/IR/Use.h"
75 #include "llvm/IR/User.h"
76 #include "llvm/IR/Value.h"
77 #include "llvm/InitializePasses.h"
78 #include "llvm/Pass.h"
79 #include "llvm/Support/BranchProbability.h"
80 #include "llvm/Support/Casting.h"
81 #include "llvm/Support/CommandLine.h"
82 #include "llvm/Support/Compiler.h"
83 #include "llvm/Support/Debug.h"
84 #include "llvm/Support/ErrorHandling.h"
85 #include "llvm/Support/raw_ostream.h"
86 #include "llvm/Transforms/Scalar.h"
87 #include "llvm/Transforms/Utils/Cloning.h"
88 #include "llvm/Transforms/Utils/LoopSimplify.h"
89 #include "llvm/Transforms/Utils/LoopUtils.h"
90 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
91 #include "llvm/Transforms/Utils/ValueMapper.h"
92 #include <algorithm>
93 #include <cassert>
94 #include <iterator>
95 #include <limits>
96 #include <utility>
97 #include <vector>
98 
99 using namespace llvm;
100 using namespace llvm::PatternMatch;
101 
102 static cl::opt<unsigned> LoopSizeCutoff("irce-loop-size-cutoff", cl::Hidden,
103                                         cl::init(64));
104 
105 static cl::opt<bool> PrintChangedLoops("irce-print-changed-loops", cl::Hidden,
106                                        cl::init(false));
107 
108 static cl::opt<bool> PrintRangeChecks("irce-print-range-checks", cl::Hidden,
109                                       cl::init(false));
110 
111 static cl::opt<bool> SkipProfitabilityChecks("irce-skip-profitability-checks",
112                                              cl::Hidden, cl::init(false));
113 
114 static cl::opt<unsigned> MinRuntimeIterations("irce-min-runtime-iterations",
115                                               cl::Hidden, cl::init(10));
116 
117 static cl::opt<bool> AllowUnsignedLatchCondition("irce-allow-unsigned-latch",
118                                                  cl::Hidden, cl::init(true));
119 
120 static cl::opt<bool> AllowNarrowLatchCondition(
121     "irce-allow-narrow-latch", cl::Hidden, cl::init(true),
122     cl::desc("If set to true, IRCE may eliminate wide range checks in loops "
123              "with narrow latch condition."));
124 
125 static const char *ClonedLoopTag = "irce.loop.clone";
126 
127 #define DEBUG_TYPE "irce"
128 
129 namespace {
130 
131 /// An inductive range check is conditional branch in a loop with
132 ///
133 ///  1. a very cold successor (i.e. the branch jumps to that successor very
134 ///     rarely)
135 ///
136 ///  and
137 ///
138 ///  2. a condition that is provably true for some contiguous range of values
139 ///     taken by the containing loop's induction variable.
140 ///
141 class InductiveRangeCheck {
142 
143   const SCEV *Begin = nullptr;
144   const SCEV *Step = nullptr;
145   const SCEV *End = nullptr;
146   Use *CheckUse = nullptr;
147 
148   static bool parseRangeCheckICmp(Loop *L, ICmpInst *ICI, ScalarEvolution &SE,
149                                   Value *&Index, Value *&Length,
150                                   bool &IsSigned);
151 
152   static void
153   extractRangeChecksFromCond(Loop *L, ScalarEvolution &SE, Use &ConditionUse,
154                              SmallVectorImpl<InductiveRangeCheck> &Checks,
155                              SmallPtrSetImpl<Value *> &Visited);
156 
157 public:
158   const SCEV *getBegin() const { return Begin; }
159   const SCEV *getStep() const { return Step; }
160   const SCEV *getEnd() const { return End; }
161 
162   void print(raw_ostream &OS) const {
163     OS << "InductiveRangeCheck:\n";
164     OS << "  Begin: ";
165     Begin->print(OS);
166     OS << "  Step: ";
167     Step->print(OS);
168     OS << "  End: ";
169     End->print(OS);
170     OS << "\n  CheckUse: ";
171     getCheckUse()->getUser()->print(OS);
172     OS << " Operand: " << getCheckUse()->getOperandNo() << "\n";
173   }
174 
175   LLVM_DUMP_METHOD
176   void dump() {
177     print(dbgs());
178   }
179 
180   Use *getCheckUse() const { return CheckUse; }
181 
182   /// Represents an signed integer range [Range.getBegin(), Range.getEnd()).  If
183   /// R.getEnd() le R.getBegin(), then R denotes the empty range.
184 
185   class Range {
186     const SCEV *Begin;
187     const SCEV *End;
188 
189   public:
190     Range(const SCEV *Begin, const SCEV *End) : Begin(Begin), End(End) {
191       assert(Begin->getType() == End->getType() && "ill-typed range!");
192     }
193 
194     Type *getType() const { return Begin->getType(); }
195     const SCEV *getBegin() const { return Begin; }
196     const SCEV *getEnd() const { return End; }
197     bool isEmpty(ScalarEvolution &SE, bool IsSigned) const {
198       if (Begin == End)
199         return true;
200       if (IsSigned)
201         return SE.isKnownPredicate(ICmpInst::ICMP_SGE, Begin, End);
202       else
203         return SE.isKnownPredicate(ICmpInst::ICMP_UGE, Begin, End);
204     }
205   };
206 
207   /// This is the value the condition of the branch needs to evaluate to for the
208   /// branch to take the hot successor (see (1) above).
209   bool getPassingDirection() { return true; }
210 
211   /// Computes a range for the induction variable (IndVar) in which the range
212   /// check is redundant and can be constant-folded away.  The induction
213   /// variable is not required to be the canonical {0,+,1} induction variable.
214   Optional<Range> computeSafeIterationSpace(ScalarEvolution &SE,
215                                             const SCEVAddRecExpr *IndVar,
216                                             bool IsLatchSigned) const;
217 
218   /// Parse out a set of inductive range checks from \p BI and append them to \p
219   /// Checks.
220   ///
221   /// NB! There may be conditions feeding into \p BI that aren't inductive range
222   /// checks, and hence don't end up in \p Checks.
223   static void
224   extractRangeChecksFromBranch(BranchInst *BI, Loop *L, ScalarEvolution &SE,
225                                BranchProbabilityInfo *BPI,
226                                SmallVectorImpl<InductiveRangeCheck> &Checks);
227 };
228 
229 struct LoopStructure;
230 
231 class InductiveRangeCheckElimination {
232   ScalarEvolution &SE;
233   BranchProbabilityInfo *BPI;
234   DominatorTree &DT;
235   LoopInfo &LI;
236 
237   using GetBFIFunc =
238       llvm::Optional<llvm::function_ref<llvm::BlockFrequencyInfo &()> >;
239   GetBFIFunc GetBFI;
240 
241   // Returns true if it is profitable to do a transform basing on estimation of
242   // number of iterations.
243   bool isProfitableToTransform(const Loop &L, LoopStructure &LS);
244 
245 public:
246   InductiveRangeCheckElimination(ScalarEvolution &SE,
247                                  BranchProbabilityInfo *BPI, DominatorTree &DT,
248                                  LoopInfo &LI, GetBFIFunc GetBFI = None)
249       : SE(SE), BPI(BPI), DT(DT), LI(LI), GetBFI(GetBFI) {}
250 
251   bool run(Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop);
252 };
253 
254 class IRCELegacyPass : public FunctionPass {
255 public:
256   static char ID;
257 
258   IRCELegacyPass() : FunctionPass(ID) {
259     initializeIRCELegacyPassPass(*PassRegistry::getPassRegistry());
260   }
261 
262   void getAnalysisUsage(AnalysisUsage &AU) const override {
263     AU.addRequired<BranchProbabilityInfoWrapperPass>();
264     AU.addRequired<DominatorTreeWrapperPass>();
265     AU.addPreserved<DominatorTreeWrapperPass>();
266     AU.addRequired<LoopInfoWrapperPass>();
267     AU.addPreserved<LoopInfoWrapperPass>();
268     AU.addRequired<ScalarEvolutionWrapperPass>();
269     AU.addPreserved<ScalarEvolutionWrapperPass>();
270   }
271 
272   bool runOnFunction(Function &F) override;
273 };
274 
275 } // end anonymous namespace
276 
277 char IRCELegacyPass::ID = 0;
278 
279 INITIALIZE_PASS_BEGIN(IRCELegacyPass, "irce",
280                       "Inductive range check elimination", false, false)
281 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
282 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
283 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
284 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
285 INITIALIZE_PASS_END(IRCELegacyPass, "irce", "Inductive range check elimination",
286                     false, false)
287 
288 /// Parse a single ICmp instruction, `ICI`, into a range check.  If `ICI` cannot
289 /// be interpreted as a range check, return false and set `Index` and `Length`
290 /// to `nullptr`.  Otherwise set `Index` to the value being range checked, and
291 /// set `Length` to the upper limit `Index` is being range checked.
292 bool
293 InductiveRangeCheck::parseRangeCheckICmp(Loop *L, ICmpInst *ICI,
294                                          ScalarEvolution &SE, Value *&Index,
295                                          Value *&Length, bool &IsSigned) {
296   auto IsLoopInvariant = [&SE, L](Value *V) {
297     return SE.isLoopInvariant(SE.getSCEV(V), L);
298   };
299 
300   ICmpInst::Predicate Pred = ICI->getPredicate();
301   Value *LHS = ICI->getOperand(0);
302   Value *RHS = ICI->getOperand(1);
303 
304   switch (Pred) {
305   default:
306     return false;
307 
308   case ICmpInst::ICMP_SLE:
309     std::swap(LHS, RHS);
310     LLVM_FALLTHROUGH;
311   case ICmpInst::ICMP_SGE:
312     IsSigned = true;
313     if (match(RHS, m_ConstantInt<0>())) {
314       Index = LHS;
315       return true; // Lower.
316     }
317     return false;
318 
319   case ICmpInst::ICMP_SLT:
320     std::swap(LHS, RHS);
321     LLVM_FALLTHROUGH;
322   case ICmpInst::ICMP_SGT:
323     IsSigned = true;
324     if (match(RHS, m_ConstantInt<-1>())) {
325       Index = LHS;
326       return true; // Lower.
327     }
328 
329     if (IsLoopInvariant(LHS)) {
330       Index = RHS;
331       Length = LHS;
332       return true; // Upper.
333     }
334     return false;
335 
336   case ICmpInst::ICMP_ULT:
337     std::swap(LHS, RHS);
338     LLVM_FALLTHROUGH;
339   case ICmpInst::ICMP_UGT:
340     IsSigned = false;
341     if (IsLoopInvariant(LHS)) {
342       Index = RHS;
343       Length = LHS;
344       return true; // Both lower and upper.
345     }
346     return false;
347   }
348 
349   llvm_unreachable("default clause returns!");
350 }
351 
352 void InductiveRangeCheck::extractRangeChecksFromCond(
353     Loop *L, ScalarEvolution &SE, Use &ConditionUse,
354     SmallVectorImpl<InductiveRangeCheck> &Checks,
355     SmallPtrSetImpl<Value *> &Visited) {
356   Value *Condition = ConditionUse.get();
357   if (!Visited.insert(Condition).second)
358     return;
359 
360   // TODO: Do the same for OR, XOR, NOT etc?
361   if (match(Condition, m_LogicalAnd(m_Value(), m_Value()))) {
362     extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(0),
363                                Checks, Visited);
364     extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(1),
365                                Checks, Visited);
366     return;
367   }
368 
369   ICmpInst *ICI = dyn_cast<ICmpInst>(Condition);
370   if (!ICI)
371     return;
372 
373   Value *Length = nullptr, *Index;
374   bool IsSigned;
375   if (!parseRangeCheckICmp(L, ICI, SE, Index, Length, IsSigned))
376     return;
377 
378   const auto *IndexAddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Index));
379   bool IsAffineIndex =
380       IndexAddRec && (IndexAddRec->getLoop() == L) && IndexAddRec->isAffine();
381 
382   if (!IsAffineIndex)
383     return;
384 
385   const SCEV *End = nullptr;
386   // We strengthen "0 <= I" to "0 <= I < INT_SMAX" and "I < L" to "0 <= I < L".
387   // We can potentially do much better here.
388   if (Length)
389     End = SE.getSCEV(Length);
390   else {
391     // So far we can only reach this point for Signed range check. This may
392     // change in future. In this case we will need to pick Unsigned max for the
393     // unsigned range check.
394     unsigned BitWidth = cast<IntegerType>(IndexAddRec->getType())->getBitWidth();
395     const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
396     End = SIntMax;
397   }
398 
399   InductiveRangeCheck IRC;
400   IRC.End = End;
401   IRC.Begin = IndexAddRec->getStart();
402   IRC.Step = IndexAddRec->getStepRecurrence(SE);
403   IRC.CheckUse = &ConditionUse;
404   Checks.push_back(IRC);
405 }
406 
407 void InductiveRangeCheck::extractRangeChecksFromBranch(
408     BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo *BPI,
409     SmallVectorImpl<InductiveRangeCheck> &Checks) {
410   if (BI->isUnconditional() || BI->getParent() == L->getLoopLatch())
411     return;
412 
413   BranchProbability LikelyTaken(15, 16);
414 
415   if (!SkipProfitabilityChecks && BPI &&
416       BPI->getEdgeProbability(BI->getParent(), (unsigned)0) < LikelyTaken)
417     return;
418 
419   SmallPtrSet<Value *, 8> Visited;
420   InductiveRangeCheck::extractRangeChecksFromCond(L, SE, BI->getOperandUse(0),
421                                                   Checks, Visited);
422 }
423 
424 // Add metadata to the loop L to disable loop optimizations. Callers need to
425 // confirm that optimizing loop L is not beneficial.
426 static void DisableAllLoopOptsOnLoop(Loop &L) {
427   // We do not care about any existing loopID related metadata for L, since we
428   // are setting all loop metadata to false.
429   LLVMContext &Context = L.getHeader()->getContext();
430   // Reserve first location for self reference to the LoopID metadata node.
431   MDNode *Dummy = MDNode::get(Context, {});
432   MDNode *DisableUnroll = MDNode::get(
433       Context, {MDString::get(Context, "llvm.loop.unroll.disable")});
434   Metadata *FalseVal =
435       ConstantAsMetadata::get(ConstantInt::get(Type::getInt1Ty(Context), 0));
436   MDNode *DisableVectorize = MDNode::get(
437       Context,
438       {MDString::get(Context, "llvm.loop.vectorize.enable"), FalseVal});
439   MDNode *DisableLICMVersioning = MDNode::get(
440       Context, {MDString::get(Context, "llvm.loop.licm_versioning.disable")});
441   MDNode *DisableDistribution= MDNode::get(
442       Context,
443       {MDString::get(Context, "llvm.loop.distribute.enable"), FalseVal});
444   MDNode *NewLoopID =
445       MDNode::get(Context, {Dummy, DisableUnroll, DisableVectorize,
446                             DisableLICMVersioning, DisableDistribution});
447   // Set operand 0 to refer to the loop id itself.
448   NewLoopID->replaceOperandWith(0, NewLoopID);
449   L.setLoopID(NewLoopID);
450 }
451 
452 namespace {
453 
454 // Keeps track of the structure of a loop.  This is similar to llvm::Loop,
455 // except that it is more lightweight and can track the state of a loop through
456 // changing and potentially invalid IR.  This structure also formalizes the
457 // kinds of loops we can deal with -- ones that have a single latch that is also
458 // an exiting block *and* have a canonical induction variable.
459 struct LoopStructure {
460   const char *Tag = "";
461 
462   BasicBlock *Header = nullptr;
463   BasicBlock *Latch = nullptr;
464 
465   // `Latch's terminator instruction is `LatchBr', and it's `LatchBrExitIdx'th
466   // successor is `LatchExit', the exit block of the loop.
467   BranchInst *LatchBr = nullptr;
468   BasicBlock *LatchExit = nullptr;
469   unsigned LatchBrExitIdx = std::numeric_limits<unsigned>::max();
470 
471   // The loop represented by this instance of LoopStructure is semantically
472   // equivalent to:
473   //
474   // intN_ty inc = IndVarIncreasing ? 1 : -1;
475   // pred_ty predicate = IndVarIncreasing ? ICMP_SLT : ICMP_SGT;
476   //
477   // for (intN_ty iv = IndVarStart; predicate(iv, LoopExitAt); iv = IndVarBase)
478   //   ... body ...
479 
480   Value *IndVarBase = nullptr;
481   Value *IndVarStart = nullptr;
482   Value *IndVarStep = nullptr;
483   Value *LoopExitAt = nullptr;
484   bool IndVarIncreasing = false;
485   bool IsSignedPredicate = true;
486 
487   LoopStructure() = default;
488 
489   template <typename M> LoopStructure map(M Map) const {
490     LoopStructure Result;
491     Result.Tag = Tag;
492     Result.Header = cast<BasicBlock>(Map(Header));
493     Result.Latch = cast<BasicBlock>(Map(Latch));
494     Result.LatchBr = cast<BranchInst>(Map(LatchBr));
495     Result.LatchExit = cast<BasicBlock>(Map(LatchExit));
496     Result.LatchBrExitIdx = LatchBrExitIdx;
497     Result.IndVarBase = Map(IndVarBase);
498     Result.IndVarStart = Map(IndVarStart);
499     Result.IndVarStep = Map(IndVarStep);
500     Result.LoopExitAt = Map(LoopExitAt);
501     Result.IndVarIncreasing = IndVarIncreasing;
502     Result.IsSignedPredicate = IsSignedPredicate;
503     return Result;
504   }
505 
506   static Optional<LoopStructure> parseLoopStructure(ScalarEvolution &, Loop &,
507                                                     const char *&);
508 };
509 
510 /// This class is used to constrain loops to run within a given iteration space.
511 /// The algorithm this class implements is given a Loop and a range [Begin,
512 /// End).  The algorithm then tries to break out a "main loop" out of the loop
513 /// it is given in a way that the "main loop" runs with the induction variable
514 /// in a subset of [Begin, End).  The algorithm emits appropriate pre and post
515 /// loops to run any remaining iterations.  The pre loop runs any iterations in
516 /// which the induction variable is < Begin, and the post loop runs any
517 /// iterations in which the induction variable is >= End.
518 class LoopConstrainer {
519   // The representation of a clone of the original loop we started out with.
520   struct ClonedLoop {
521     // The cloned blocks
522     std::vector<BasicBlock *> Blocks;
523 
524     // `Map` maps values in the clonee into values in the cloned version
525     ValueToValueMapTy Map;
526 
527     // An instance of `LoopStructure` for the cloned loop
528     LoopStructure Structure;
529   };
530 
531   // Result of rewriting the range of a loop.  See changeIterationSpaceEnd for
532   // more details on what these fields mean.
533   struct RewrittenRangeInfo {
534     BasicBlock *PseudoExit = nullptr;
535     BasicBlock *ExitSelector = nullptr;
536     std::vector<PHINode *> PHIValuesAtPseudoExit;
537     PHINode *IndVarEnd = nullptr;
538 
539     RewrittenRangeInfo() = default;
540   };
541 
542   // Calculated subranges we restrict the iteration space of the main loop to.
543   // See the implementation of `calculateSubRanges' for more details on how
544   // these fields are computed.  `LowLimit` is None if there is no restriction
545   // on low end of the restricted iteration space of the main loop.  `HighLimit`
546   // is None if there is no restriction on high end of the restricted iteration
547   // space of the main loop.
548 
549   struct SubRanges {
550     Optional<const SCEV *> LowLimit;
551     Optional<const SCEV *> HighLimit;
552   };
553 
554   // Compute a safe set of limits for the main loop to run in -- effectively the
555   // intersection of `Range' and the iteration space of the original loop.
556   // Return None if unable to compute the set of subranges.
557   Optional<SubRanges> calculateSubRanges(bool IsSignedPredicate) const;
558 
559   // Clone `OriginalLoop' and return the result in CLResult.  The IR after
560   // running `cloneLoop' is well formed except for the PHI nodes in CLResult --
561   // the PHI nodes say that there is an incoming edge from `OriginalPreheader`
562   // but there is no such edge.
563   void cloneLoop(ClonedLoop &CLResult, const char *Tag) const;
564 
565   // Create the appropriate loop structure needed to describe a cloned copy of
566   // `Original`.  The clone is described by `VM`.
567   Loop *createClonedLoopStructure(Loop *Original, Loop *Parent,
568                                   ValueToValueMapTy &VM, bool IsSubloop);
569 
570   // Rewrite the iteration space of the loop denoted by (LS, Preheader). The
571   // iteration space of the rewritten loop ends at ExitLoopAt.  The start of the
572   // iteration space is not changed.  `ExitLoopAt' is assumed to be slt
573   // `OriginalHeaderCount'.
574   //
575   // If there are iterations left to execute, control is made to jump to
576   // `ContinuationBlock', otherwise they take the normal loop exit.  The
577   // returned `RewrittenRangeInfo' object is populated as follows:
578   //
579   //  .PseudoExit is a basic block that unconditionally branches to
580   //      `ContinuationBlock'.
581   //
582   //  .ExitSelector is a basic block that decides, on exit from the loop,
583   //      whether to branch to the "true" exit or to `PseudoExit'.
584   //
585   //  .PHIValuesAtPseudoExit are PHINodes in `PseudoExit' that compute the value
586   //      for each PHINode in the loop header on taking the pseudo exit.
587   //
588   // After changeIterationSpaceEnd, `Preheader' is no longer a legitimate
589   // preheader because it is made to branch to the loop header only
590   // conditionally.
591   RewrittenRangeInfo
592   changeIterationSpaceEnd(const LoopStructure &LS, BasicBlock *Preheader,
593                           Value *ExitLoopAt,
594                           BasicBlock *ContinuationBlock) const;
595 
596   // The loop denoted by `LS' has `OldPreheader' as its preheader.  This
597   // function creates a new preheader for `LS' and returns it.
598   BasicBlock *createPreheader(const LoopStructure &LS, BasicBlock *OldPreheader,
599                               const char *Tag) const;
600 
601   // `ContinuationBlockAndPreheader' was the continuation block for some call to
602   // `changeIterationSpaceEnd' and is the preheader to the loop denoted by `LS'.
603   // This function rewrites the PHI nodes in `LS.Header' to start with the
604   // correct value.
605   void rewriteIncomingValuesForPHIs(
606       LoopStructure &LS, BasicBlock *ContinuationBlockAndPreheader,
607       const LoopConstrainer::RewrittenRangeInfo &RRI) const;
608 
609   // Even though we do not preserve any passes at this time, we at least need to
610   // keep the parent loop structure consistent.  The `LPPassManager' seems to
611   // verify this after running a loop pass.  This function adds the list of
612   // blocks denoted by BBs to this loops parent loop if required.
613   void addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs);
614 
615   // Some global state.
616   Function &F;
617   LLVMContext &Ctx;
618   ScalarEvolution &SE;
619   DominatorTree &DT;
620   LoopInfo &LI;
621   function_ref<void(Loop *, bool)> LPMAddNewLoop;
622 
623   // Information about the original loop we started out with.
624   Loop &OriginalLoop;
625 
626   const SCEV *LatchTakenCount = nullptr;
627   BasicBlock *OriginalPreheader = nullptr;
628 
629   // The preheader of the main loop.  This may or may not be different from
630   // `OriginalPreheader'.
631   BasicBlock *MainLoopPreheader = nullptr;
632 
633   // The range we need to run the main loop in.
634   InductiveRangeCheck::Range Range;
635 
636   // The structure of the main loop (see comment at the beginning of this class
637   // for a definition)
638   LoopStructure MainLoopStructure;
639 
640 public:
641   LoopConstrainer(Loop &L, LoopInfo &LI,
642                   function_ref<void(Loop *, bool)> LPMAddNewLoop,
643                   const LoopStructure &LS, ScalarEvolution &SE,
644                   DominatorTree &DT, InductiveRangeCheck::Range R)
645       : F(*L.getHeader()->getParent()), Ctx(L.getHeader()->getContext()),
646         SE(SE), DT(DT), LI(LI), LPMAddNewLoop(LPMAddNewLoop), OriginalLoop(L),
647         Range(R), MainLoopStructure(LS) {}
648 
649   // Entry point for the algorithm.  Returns true on success.
650   bool run();
651 };
652 
653 } // end anonymous namespace
654 
655 /// Given a loop with an deccreasing induction variable, is it possible to
656 /// safely calculate the bounds of a new loop using the given Predicate.
657 static bool isSafeDecreasingBound(const SCEV *Start,
658                                   const SCEV *BoundSCEV, const SCEV *Step,
659                                   ICmpInst::Predicate Pred,
660                                   unsigned LatchBrExitIdx,
661                                   Loop *L, ScalarEvolution &SE) {
662   if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT &&
663       Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT)
664     return false;
665 
666   if (!SE.isAvailableAtLoopEntry(BoundSCEV, L))
667     return false;
668 
669   assert(SE.isKnownNegative(Step) && "expecting negative step");
670 
671   LLVM_DEBUG(dbgs() << "irce: isSafeDecreasingBound with:\n");
672   LLVM_DEBUG(dbgs() << "irce: Start: " << *Start << "\n");
673   LLVM_DEBUG(dbgs() << "irce: Step: " << *Step << "\n");
674   LLVM_DEBUG(dbgs() << "irce: BoundSCEV: " << *BoundSCEV << "\n");
675   LLVM_DEBUG(dbgs() << "irce: Pred: " << ICmpInst::getPredicateName(Pred)
676                     << "\n");
677   LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n");
678 
679   bool IsSigned = ICmpInst::isSigned(Pred);
680   // The predicate that we need to check that the induction variable lies
681   // within bounds.
682   ICmpInst::Predicate BoundPred =
683     IsSigned ? CmpInst::ICMP_SGT : CmpInst::ICMP_UGT;
684 
685   if (LatchBrExitIdx == 1)
686     return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV);
687 
688   assert(LatchBrExitIdx == 0 &&
689          "LatchBrExitIdx should be either 0 or 1");
690 
691   const SCEV *StepPlusOne = SE.getAddExpr(Step, SE.getOne(Step->getType()));
692   unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
693   APInt Min = IsSigned ? APInt::getSignedMinValue(BitWidth) :
694     APInt::getMinValue(BitWidth);
695   const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Min), StepPlusOne);
696 
697   const SCEV *MinusOne =
698     SE.getMinusSCEV(BoundSCEV, SE.getOne(BoundSCEV->getType()));
699 
700   return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, MinusOne) &&
701          SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit);
702 
703 }
704 
705 /// Given a loop with an increasing induction variable, is it possible to
706 /// safely calculate the bounds of a new loop using the given Predicate.
707 static bool isSafeIncreasingBound(const SCEV *Start,
708                                   const SCEV *BoundSCEV, const SCEV *Step,
709                                   ICmpInst::Predicate Pred,
710                                   unsigned LatchBrExitIdx,
711                                   Loop *L, ScalarEvolution &SE) {
712   if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT &&
713       Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT)
714     return false;
715 
716   if (!SE.isAvailableAtLoopEntry(BoundSCEV, L))
717     return false;
718 
719   LLVM_DEBUG(dbgs() << "irce: isSafeIncreasingBound with:\n");
720   LLVM_DEBUG(dbgs() << "irce: Start: " << *Start << "\n");
721   LLVM_DEBUG(dbgs() << "irce: Step: " << *Step << "\n");
722   LLVM_DEBUG(dbgs() << "irce: BoundSCEV: " << *BoundSCEV << "\n");
723   LLVM_DEBUG(dbgs() << "irce: Pred: " << ICmpInst::getPredicateName(Pred)
724                     << "\n");
725   LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n");
726 
727   bool IsSigned = ICmpInst::isSigned(Pred);
728   // The predicate that we need to check that the induction variable lies
729   // within bounds.
730   ICmpInst::Predicate BoundPred =
731       IsSigned ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
732 
733   if (LatchBrExitIdx == 1)
734     return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV);
735 
736   assert(LatchBrExitIdx == 0 && "LatchBrExitIdx should be 0 or 1");
737 
738   const SCEV *StepMinusOne =
739     SE.getMinusSCEV(Step, SE.getOne(Step->getType()));
740   unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
741   APInt Max = IsSigned ? APInt::getSignedMaxValue(BitWidth) :
742     APInt::getMaxValue(BitWidth);
743   const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Max), StepMinusOne);
744 
745   return (SE.isLoopEntryGuardedByCond(L, BoundPred, Start,
746                                       SE.getAddExpr(BoundSCEV, Step)) &&
747           SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit));
748 }
749 
750 Optional<LoopStructure>
751 LoopStructure::parseLoopStructure(ScalarEvolution &SE, Loop &L,
752                                   const char *&FailureReason) {
753   if (!L.isLoopSimplifyForm()) {
754     FailureReason = "loop not in LoopSimplify form";
755     return None;
756   }
757 
758   BasicBlock *Latch = L.getLoopLatch();
759   assert(Latch && "Simplified loops only have one latch!");
760 
761   if (Latch->getTerminator()->getMetadata(ClonedLoopTag)) {
762     FailureReason = "loop has already been cloned";
763     return None;
764   }
765 
766   if (!L.isLoopExiting(Latch)) {
767     FailureReason = "no loop latch";
768     return None;
769   }
770 
771   BasicBlock *Header = L.getHeader();
772   BasicBlock *Preheader = L.getLoopPreheader();
773   if (!Preheader) {
774     FailureReason = "no preheader";
775     return None;
776   }
777 
778   BranchInst *LatchBr = dyn_cast<BranchInst>(Latch->getTerminator());
779   if (!LatchBr || LatchBr->isUnconditional()) {
780     FailureReason = "latch terminator not conditional branch";
781     return None;
782   }
783 
784   unsigned LatchBrExitIdx = LatchBr->getSuccessor(0) == Header ? 1 : 0;
785 
786   ICmpInst *ICI = dyn_cast<ICmpInst>(LatchBr->getCondition());
787   if (!ICI || !isa<IntegerType>(ICI->getOperand(0)->getType())) {
788     FailureReason = "latch terminator branch not conditional on integral icmp";
789     return None;
790   }
791 
792   const SCEV *LatchCount = SE.getExitCount(&L, Latch);
793   if (isa<SCEVCouldNotCompute>(LatchCount)) {
794     FailureReason = "could not compute latch count";
795     return None;
796   }
797 
798   ICmpInst::Predicate Pred = ICI->getPredicate();
799   Value *LeftValue = ICI->getOperand(0);
800   const SCEV *LeftSCEV = SE.getSCEV(LeftValue);
801   IntegerType *IndVarTy = cast<IntegerType>(LeftValue->getType());
802 
803   Value *RightValue = ICI->getOperand(1);
804   const SCEV *RightSCEV = SE.getSCEV(RightValue);
805 
806   // We canonicalize `ICI` such that `LeftSCEV` is an add recurrence.
807   if (!isa<SCEVAddRecExpr>(LeftSCEV)) {
808     if (isa<SCEVAddRecExpr>(RightSCEV)) {
809       std::swap(LeftSCEV, RightSCEV);
810       std::swap(LeftValue, RightValue);
811       Pred = ICmpInst::getSwappedPredicate(Pred);
812     } else {
813       FailureReason = "no add recurrences in the icmp";
814       return None;
815     }
816   }
817 
818   auto HasNoSignedWrap = [&](const SCEVAddRecExpr *AR) {
819     if (AR->getNoWrapFlags(SCEV::FlagNSW))
820       return true;
821 
822     IntegerType *Ty = cast<IntegerType>(AR->getType());
823     IntegerType *WideTy =
824         IntegerType::get(Ty->getContext(), Ty->getBitWidth() * 2);
825 
826     const SCEVAddRecExpr *ExtendAfterOp =
827         dyn_cast<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
828     if (ExtendAfterOp) {
829       const SCEV *ExtendedStart = SE.getSignExtendExpr(AR->getStart(), WideTy);
830       const SCEV *ExtendedStep =
831           SE.getSignExtendExpr(AR->getStepRecurrence(SE), WideTy);
832 
833       bool NoSignedWrap = ExtendAfterOp->getStart() == ExtendedStart &&
834                           ExtendAfterOp->getStepRecurrence(SE) == ExtendedStep;
835 
836       if (NoSignedWrap)
837         return true;
838     }
839 
840     // We may have proved this when computing the sign extension above.
841     return AR->getNoWrapFlags(SCEV::FlagNSW) != SCEV::FlagAnyWrap;
842   };
843 
844   // `ICI` is interpreted as taking the backedge if the *next* value of the
845   // induction variable satisfies some constraint.
846 
847   const SCEVAddRecExpr *IndVarBase = cast<SCEVAddRecExpr>(LeftSCEV);
848   if (!IndVarBase->isAffine()) {
849     FailureReason = "LHS in icmp not induction variable";
850     return None;
851   }
852   const SCEV* StepRec = IndVarBase->getStepRecurrence(SE);
853   if (!isa<SCEVConstant>(StepRec)) {
854     FailureReason = "LHS in icmp not induction variable";
855     return None;
856   }
857   ConstantInt *StepCI = cast<SCEVConstant>(StepRec)->getValue();
858 
859   if (ICI->isEquality() && !HasNoSignedWrap(IndVarBase)) {
860     FailureReason = "LHS in icmp needs nsw for equality predicates";
861     return None;
862   }
863 
864   assert(!StepCI->isZero() && "Zero step?");
865   bool IsIncreasing = !StepCI->isNegative();
866   bool IsSignedPredicate;
867   const SCEV *StartNext = IndVarBase->getStart();
868   const SCEV *Addend = SE.getNegativeSCEV(IndVarBase->getStepRecurrence(SE));
869   const SCEV *IndVarStart = SE.getAddExpr(StartNext, Addend);
870   const SCEV *Step = SE.getSCEV(StepCI);
871 
872   const SCEV *FixedRightSCEV = nullptr;
873 
874   // If RightValue resides within loop (but still being loop invariant),
875   // regenerate it as preheader.
876   if (auto *I = dyn_cast<Instruction>(RightValue))
877     if (L.contains(I->getParent()))
878       FixedRightSCEV = RightSCEV;
879 
880   if (IsIncreasing) {
881     bool DecreasedRightValueByOne = false;
882     if (StepCI->isOne()) {
883       // Try to turn eq/ne predicates to those we can work with.
884       if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
885         // while (++i != len) {         while (++i < len) {
886         //   ...                 --->     ...
887         // }                            }
888         // If both parts are known non-negative, it is profitable to use
889         // unsigned comparison in increasing loop. This allows us to make the
890         // comparison check against "RightSCEV + 1" more optimistic.
891         if (isKnownNonNegativeInLoop(IndVarStart, &L, SE) &&
892             isKnownNonNegativeInLoop(RightSCEV, &L, SE))
893           Pred = ICmpInst::ICMP_ULT;
894         else
895           Pred = ICmpInst::ICMP_SLT;
896       else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) {
897         // while (true) {               while (true) {
898         //   if (++i == len)     --->     if (++i > len - 1)
899         //     break;                       break;
900         //   ...                          ...
901         // }                            }
902         if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) &&
903             cannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/false)) {
904           Pred = ICmpInst::ICMP_UGT;
905           RightSCEV = SE.getMinusSCEV(RightSCEV,
906                                       SE.getOne(RightSCEV->getType()));
907           DecreasedRightValueByOne = true;
908         } else if (cannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/true)) {
909           Pred = ICmpInst::ICMP_SGT;
910           RightSCEV = SE.getMinusSCEV(RightSCEV,
911                                       SE.getOne(RightSCEV->getType()));
912           DecreasedRightValueByOne = true;
913         }
914       }
915     }
916 
917     bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
918     bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
919     bool FoundExpectedPred =
920         (LTPred && LatchBrExitIdx == 1) || (GTPred && LatchBrExitIdx == 0);
921 
922     if (!FoundExpectedPred) {
923       FailureReason = "expected icmp slt semantically, found something else";
924       return None;
925     }
926 
927     IsSignedPredicate = ICmpInst::isSigned(Pred);
928     if (!IsSignedPredicate && !AllowUnsignedLatchCondition) {
929       FailureReason = "unsigned latch conditions are explicitly prohibited";
930       return None;
931     }
932 
933     if (!isSafeIncreasingBound(IndVarStart, RightSCEV, Step, Pred,
934                                LatchBrExitIdx, &L, SE)) {
935       FailureReason = "Unsafe loop bounds";
936       return None;
937     }
938     if (LatchBrExitIdx == 0) {
939       // We need to increase the right value unless we have already decreased
940       // it virtually when we replaced EQ with SGT.
941       if (!DecreasedRightValueByOne)
942         FixedRightSCEV =
943             SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
944     } else {
945       assert(!DecreasedRightValueByOne &&
946              "Right value can be decreased only for LatchBrExitIdx == 0!");
947     }
948   } else {
949     bool IncreasedRightValueByOne = false;
950     if (StepCI->isMinusOne()) {
951       // Try to turn eq/ne predicates to those we can work with.
952       if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
953         // while (--i != len) {         while (--i > len) {
954         //   ...                 --->     ...
955         // }                            }
956         // We intentionally don't turn the predicate into UGT even if we know
957         // that both operands are non-negative, because it will only pessimize
958         // our check against "RightSCEV - 1".
959         Pred = ICmpInst::ICMP_SGT;
960       else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) {
961         // while (true) {               while (true) {
962         //   if (--i == len)     --->     if (--i < len + 1)
963         //     break;                       break;
964         //   ...                          ...
965         // }                            }
966         if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) &&
967             cannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ false)) {
968           Pred = ICmpInst::ICMP_ULT;
969           RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
970           IncreasedRightValueByOne = true;
971         } else if (cannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ true)) {
972           Pred = ICmpInst::ICMP_SLT;
973           RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
974           IncreasedRightValueByOne = true;
975         }
976       }
977     }
978 
979     bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
980     bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
981 
982     bool FoundExpectedPred =
983         (GTPred && LatchBrExitIdx == 1) || (LTPred && LatchBrExitIdx == 0);
984 
985     if (!FoundExpectedPred) {
986       FailureReason = "expected icmp sgt semantically, found something else";
987       return None;
988     }
989 
990     IsSignedPredicate =
991         Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGT;
992 
993     if (!IsSignedPredicate && !AllowUnsignedLatchCondition) {
994       FailureReason = "unsigned latch conditions are explicitly prohibited";
995       return None;
996     }
997 
998     if (!isSafeDecreasingBound(IndVarStart, RightSCEV, Step, Pred,
999                                LatchBrExitIdx, &L, SE)) {
1000       FailureReason = "Unsafe bounds";
1001       return None;
1002     }
1003 
1004     if (LatchBrExitIdx == 0) {
1005       // We need to decrease the right value unless we have already increased
1006       // it virtually when we replaced EQ with SLT.
1007       if (!IncreasedRightValueByOne)
1008         FixedRightSCEV =
1009             SE.getMinusSCEV(RightSCEV, SE.getOne(RightSCEV->getType()));
1010     } else {
1011       assert(!IncreasedRightValueByOne &&
1012              "Right value can be increased only for LatchBrExitIdx == 0!");
1013     }
1014   }
1015   BasicBlock *LatchExit = LatchBr->getSuccessor(LatchBrExitIdx);
1016 
1017   assert(SE.getLoopDisposition(LatchCount, &L) ==
1018              ScalarEvolution::LoopInvariant &&
1019          "loop variant exit count doesn't make sense!");
1020 
1021   assert(!L.contains(LatchExit) && "expected an exit block!");
1022   const DataLayout &DL = Preheader->getModule()->getDataLayout();
1023   SCEVExpander Expander(SE, DL, "irce");
1024   Instruction *Ins = Preheader->getTerminator();
1025 
1026   if (FixedRightSCEV)
1027     RightValue =
1028         Expander.expandCodeFor(FixedRightSCEV, FixedRightSCEV->getType(), Ins);
1029 
1030   Value *IndVarStartV = Expander.expandCodeFor(IndVarStart, IndVarTy, Ins);
1031   IndVarStartV->setName("indvar.start");
1032 
1033   LoopStructure Result;
1034 
1035   Result.Tag = "main";
1036   Result.Header = Header;
1037   Result.Latch = Latch;
1038   Result.LatchBr = LatchBr;
1039   Result.LatchExit = LatchExit;
1040   Result.LatchBrExitIdx = LatchBrExitIdx;
1041   Result.IndVarStart = IndVarStartV;
1042   Result.IndVarStep = StepCI;
1043   Result.IndVarBase = LeftValue;
1044   Result.IndVarIncreasing = IsIncreasing;
1045   Result.LoopExitAt = RightValue;
1046   Result.IsSignedPredicate = IsSignedPredicate;
1047 
1048   FailureReason = nullptr;
1049 
1050   return Result;
1051 }
1052 
1053 /// If the type of \p S matches with \p Ty, return \p S. Otherwise, return
1054 /// signed or unsigned extension of \p S to type \p Ty.
1055 static const SCEV *NoopOrExtend(const SCEV *S, Type *Ty, ScalarEvolution &SE,
1056                                 bool Signed) {
1057   return Signed ? SE.getNoopOrSignExtend(S, Ty) : SE.getNoopOrZeroExtend(S, Ty);
1058 }
1059 
1060 Optional<LoopConstrainer::SubRanges>
1061 LoopConstrainer::calculateSubRanges(bool IsSignedPredicate) const {
1062   IntegerType *Ty = cast<IntegerType>(LatchTakenCount->getType());
1063 
1064   auto *RTy = cast<IntegerType>(Range.getType());
1065 
1066   // We only support wide range checks and narrow latches.
1067   if (!AllowNarrowLatchCondition && RTy != Ty)
1068     return None;
1069   if (RTy->getBitWidth() < Ty->getBitWidth())
1070     return None;
1071 
1072   LoopConstrainer::SubRanges Result;
1073 
1074   // I think we can be more aggressive here and make this nuw / nsw if the
1075   // addition that feeds into the icmp for the latch's terminating branch is nuw
1076   // / nsw.  In any case, a wrapping 2's complement addition is safe.
1077   const SCEV *Start = NoopOrExtend(SE.getSCEV(MainLoopStructure.IndVarStart),
1078                                    RTy, SE, IsSignedPredicate);
1079   const SCEV *End = NoopOrExtend(SE.getSCEV(MainLoopStructure.LoopExitAt), RTy,
1080                                  SE, IsSignedPredicate);
1081 
1082   bool Increasing = MainLoopStructure.IndVarIncreasing;
1083 
1084   // We compute `Smallest` and `Greatest` such that [Smallest, Greatest), or
1085   // [Smallest, GreatestSeen] is the range of values the induction variable
1086   // takes.
1087 
1088   const SCEV *Smallest = nullptr, *Greatest = nullptr, *GreatestSeen = nullptr;
1089 
1090   const SCEV *One = SE.getOne(RTy);
1091   if (Increasing) {
1092     Smallest = Start;
1093     Greatest = End;
1094     // No overflow, because the range [Smallest, GreatestSeen] is not empty.
1095     GreatestSeen = SE.getMinusSCEV(End, One);
1096   } else {
1097     // These two computations may sign-overflow.  Here is why that is okay:
1098     //
1099     // We know that the induction variable does not sign-overflow on any
1100     // iteration except the last one, and it starts at `Start` and ends at
1101     // `End`, decrementing by one every time.
1102     //
1103     //  * if `Smallest` sign-overflows we know `End` is `INT_SMAX`. Since the
1104     //    induction variable is decreasing we know that that the smallest value
1105     //    the loop body is actually executed with is `INT_SMIN` == `Smallest`.
1106     //
1107     //  * if `Greatest` sign-overflows, we know it can only be `INT_SMIN`.  In
1108     //    that case, `Clamp` will always return `Smallest` and
1109     //    [`Result.LowLimit`, `Result.HighLimit`) = [`Smallest`, `Smallest`)
1110     //    will be an empty range.  Returning an empty range is always safe.
1111 
1112     Smallest = SE.getAddExpr(End, One);
1113     Greatest = SE.getAddExpr(Start, One);
1114     GreatestSeen = Start;
1115   }
1116 
1117   auto Clamp = [this, Smallest, Greatest, IsSignedPredicate](const SCEV *S) {
1118     return IsSignedPredicate
1119                ? SE.getSMaxExpr(Smallest, SE.getSMinExpr(Greatest, S))
1120                : SE.getUMaxExpr(Smallest, SE.getUMinExpr(Greatest, S));
1121   };
1122 
1123   // In some cases we can prove that we don't need a pre or post loop.
1124   ICmpInst::Predicate PredLE =
1125       IsSignedPredicate ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1126   ICmpInst::Predicate PredLT =
1127       IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1128 
1129   bool ProvablyNoPreloop =
1130       SE.isKnownPredicate(PredLE, Range.getBegin(), Smallest);
1131   if (!ProvablyNoPreloop)
1132     Result.LowLimit = Clamp(Range.getBegin());
1133 
1134   bool ProvablyNoPostLoop =
1135       SE.isKnownPredicate(PredLT, GreatestSeen, Range.getEnd());
1136   if (!ProvablyNoPostLoop)
1137     Result.HighLimit = Clamp(Range.getEnd());
1138 
1139   return Result;
1140 }
1141 
1142 void LoopConstrainer::cloneLoop(LoopConstrainer::ClonedLoop &Result,
1143                                 const char *Tag) const {
1144   for (BasicBlock *BB : OriginalLoop.getBlocks()) {
1145     BasicBlock *Clone = CloneBasicBlock(BB, Result.Map, Twine(".") + Tag, &F);
1146     Result.Blocks.push_back(Clone);
1147     Result.Map[BB] = Clone;
1148   }
1149 
1150   auto GetClonedValue = [&Result](Value *V) {
1151     assert(V && "null values not in domain!");
1152     auto It = Result.Map.find(V);
1153     if (It == Result.Map.end())
1154       return V;
1155     return static_cast<Value *>(It->second);
1156   };
1157 
1158   auto *ClonedLatch =
1159       cast<BasicBlock>(GetClonedValue(OriginalLoop.getLoopLatch()));
1160   ClonedLatch->getTerminator()->setMetadata(ClonedLoopTag,
1161                                             MDNode::get(Ctx, {}));
1162 
1163   Result.Structure = MainLoopStructure.map(GetClonedValue);
1164   Result.Structure.Tag = Tag;
1165 
1166   for (unsigned i = 0, e = Result.Blocks.size(); i != e; ++i) {
1167     BasicBlock *ClonedBB = Result.Blocks[i];
1168     BasicBlock *OriginalBB = OriginalLoop.getBlocks()[i];
1169 
1170     assert(Result.Map[OriginalBB] == ClonedBB && "invariant!");
1171 
1172     for (Instruction &I : *ClonedBB)
1173       RemapInstruction(&I, Result.Map,
1174                        RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1175 
1176     // Exit blocks will now have one more predecessor and their PHI nodes need
1177     // to be edited to reflect that.  No phi nodes need to be introduced because
1178     // the loop is in LCSSA.
1179 
1180     for (auto *SBB : successors(OriginalBB)) {
1181       if (OriginalLoop.contains(SBB))
1182         continue; // not an exit block
1183 
1184       for (PHINode &PN : SBB->phis()) {
1185         Value *OldIncoming = PN.getIncomingValueForBlock(OriginalBB);
1186         PN.addIncoming(GetClonedValue(OldIncoming), ClonedBB);
1187       }
1188     }
1189   }
1190 }
1191 
1192 LoopConstrainer::RewrittenRangeInfo LoopConstrainer::changeIterationSpaceEnd(
1193     const LoopStructure &LS, BasicBlock *Preheader, Value *ExitSubloopAt,
1194     BasicBlock *ContinuationBlock) const {
1195   // We start with a loop with a single latch:
1196   //
1197   //    +--------------------+
1198   //    |                    |
1199   //    |     preheader      |
1200   //    |                    |
1201   //    +--------+-----------+
1202   //             |      ----------------\
1203   //             |     /                |
1204   //    +--------v----v------+          |
1205   //    |                    |          |
1206   //    |      header        |          |
1207   //    |                    |          |
1208   //    +--------------------+          |
1209   //                                    |
1210   //            .....                   |
1211   //                                    |
1212   //    +--------------------+          |
1213   //    |                    |          |
1214   //    |       latch        >----------/
1215   //    |                    |
1216   //    +-------v------------+
1217   //            |
1218   //            |
1219   //            |   +--------------------+
1220   //            |   |                    |
1221   //            +--->   original exit    |
1222   //                |                    |
1223   //                +--------------------+
1224   //
1225   // We change the control flow to look like
1226   //
1227   //
1228   //    +--------------------+
1229   //    |                    |
1230   //    |     preheader      >-------------------------+
1231   //    |                    |                         |
1232   //    +--------v-----------+                         |
1233   //             |    /-------------+                  |
1234   //             |   /              |                  |
1235   //    +--------v--v--------+      |                  |
1236   //    |                    |      |                  |
1237   //    |      header        |      |   +--------+     |
1238   //    |                    |      |   |        |     |
1239   //    +--------------------+      |   |  +-----v-----v-----------+
1240   //                                |   |  |                       |
1241   //                                |   |  |     .pseudo.exit      |
1242   //                                |   |  |                       |
1243   //                                |   |  +-----------v-----------+
1244   //                                |   |              |
1245   //            .....               |   |              |
1246   //                                |   |     +--------v-------------+
1247   //    +--------------------+      |   |     |                      |
1248   //    |                    |      |   |     |   ContinuationBlock  |
1249   //    |       latch        >------+   |     |                      |
1250   //    |                    |          |     +----------------------+
1251   //    +---------v----------+          |
1252   //              |                     |
1253   //              |                     |
1254   //              |     +---------------^-----+
1255   //              |     |                     |
1256   //              +----->    .exit.selector   |
1257   //                    |                     |
1258   //                    +----------v----------+
1259   //                               |
1260   //     +--------------------+    |
1261   //     |                    |    |
1262   //     |   original exit    <----+
1263   //     |                    |
1264   //     +--------------------+
1265 
1266   RewrittenRangeInfo RRI;
1267 
1268   BasicBlock *BBInsertLocation = LS.Latch->getNextNode();
1269   RRI.ExitSelector = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".exit.selector",
1270                                         &F, BBInsertLocation);
1271   RRI.PseudoExit = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".pseudo.exit", &F,
1272                                       BBInsertLocation);
1273 
1274   BranchInst *PreheaderJump = cast<BranchInst>(Preheader->getTerminator());
1275   bool Increasing = LS.IndVarIncreasing;
1276   bool IsSignedPredicate = LS.IsSignedPredicate;
1277 
1278   IRBuilder<> B(PreheaderJump);
1279   auto *RangeTy = Range.getBegin()->getType();
1280   auto NoopOrExt = [&](Value *V) {
1281     if (V->getType() == RangeTy)
1282       return V;
1283     return IsSignedPredicate ? B.CreateSExt(V, RangeTy, "wide." + V->getName())
1284                              : B.CreateZExt(V, RangeTy, "wide." + V->getName());
1285   };
1286 
1287   // EnterLoopCond - is it okay to start executing this `LS'?
1288   Value *EnterLoopCond = nullptr;
1289   auto Pred =
1290       Increasing
1291           ? (IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT)
1292           : (IsSignedPredicate ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
1293   Value *IndVarStart = NoopOrExt(LS.IndVarStart);
1294   EnterLoopCond = B.CreateICmp(Pred, IndVarStart, ExitSubloopAt);
1295 
1296   B.CreateCondBr(EnterLoopCond, LS.Header, RRI.PseudoExit);
1297   PreheaderJump->eraseFromParent();
1298 
1299   LS.LatchBr->setSuccessor(LS.LatchBrExitIdx, RRI.ExitSelector);
1300   B.SetInsertPoint(LS.LatchBr);
1301   Value *IndVarBase = NoopOrExt(LS.IndVarBase);
1302   Value *TakeBackedgeLoopCond = B.CreateICmp(Pred, IndVarBase, ExitSubloopAt);
1303 
1304   Value *CondForBranch = LS.LatchBrExitIdx == 1
1305                              ? TakeBackedgeLoopCond
1306                              : B.CreateNot(TakeBackedgeLoopCond);
1307 
1308   LS.LatchBr->setCondition(CondForBranch);
1309 
1310   B.SetInsertPoint(RRI.ExitSelector);
1311 
1312   // IterationsLeft - are there any more iterations left, given the original
1313   // upper bound on the induction variable?  If not, we branch to the "real"
1314   // exit.
1315   Value *LoopExitAt = NoopOrExt(LS.LoopExitAt);
1316   Value *IterationsLeft = B.CreateICmp(Pred, IndVarBase, LoopExitAt);
1317   B.CreateCondBr(IterationsLeft, RRI.PseudoExit, LS.LatchExit);
1318 
1319   BranchInst *BranchToContinuation =
1320       BranchInst::Create(ContinuationBlock, RRI.PseudoExit);
1321 
1322   // We emit PHI nodes into `RRI.PseudoExit' that compute the "latest" value of
1323   // each of the PHI nodes in the loop header.  This feeds into the initial
1324   // value of the same PHI nodes if/when we continue execution.
1325   for (PHINode &PN : LS.Header->phis()) {
1326     PHINode *NewPHI = PHINode::Create(PN.getType(), 2, PN.getName() + ".copy",
1327                                       BranchToContinuation);
1328 
1329     NewPHI->addIncoming(PN.getIncomingValueForBlock(Preheader), Preheader);
1330     NewPHI->addIncoming(PN.getIncomingValueForBlock(LS.Latch),
1331                         RRI.ExitSelector);
1332     RRI.PHIValuesAtPseudoExit.push_back(NewPHI);
1333   }
1334 
1335   RRI.IndVarEnd = PHINode::Create(IndVarBase->getType(), 2, "indvar.end",
1336                                   BranchToContinuation);
1337   RRI.IndVarEnd->addIncoming(IndVarStart, Preheader);
1338   RRI.IndVarEnd->addIncoming(IndVarBase, RRI.ExitSelector);
1339 
1340   // The latch exit now has a branch from `RRI.ExitSelector' instead of
1341   // `LS.Latch'.  The PHI nodes need to be updated to reflect that.
1342   LS.LatchExit->replacePhiUsesWith(LS.Latch, RRI.ExitSelector);
1343 
1344   return RRI;
1345 }
1346 
1347 void LoopConstrainer::rewriteIncomingValuesForPHIs(
1348     LoopStructure &LS, BasicBlock *ContinuationBlock,
1349     const LoopConstrainer::RewrittenRangeInfo &RRI) const {
1350   unsigned PHIIndex = 0;
1351   for (PHINode &PN : LS.Header->phis())
1352     PN.setIncomingValueForBlock(ContinuationBlock,
1353                                 RRI.PHIValuesAtPseudoExit[PHIIndex++]);
1354 
1355   LS.IndVarStart = RRI.IndVarEnd;
1356 }
1357 
1358 BasicBlock *LoopConstrainer::createPreheader(const LoopStructure &LS,
1359                                              BasicBlock *OldPreheader,
1360                                              const char *Tag) const {
1361   BasicBlock *Preheader = BasicBlock::Create(Ctx, Tag, &F, LS.Header);
1362   BranchInst::Create(LS.Header, Preheader);
1363 
1364   LS.Header->replacePhiUsesWith(OldPreheader, Preheader);
1365 
1366   return Preheader;
1367 }
1368 
1369 void LoopConstrainer::addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs) {
1370   Loop *ParentLoop = OriginalLoop.getParentLoop();
1371   if (!ParentLoop)
1372     return;
1373 
1374   for (BasicBlock *BB : BBs)
1375     ParentLoop->addBasicBlockToLoop(BB, LI);
1376 }
1377 
1378 Loop *LoopConstrainer::createClonedLoopStructure(Loop *Original, Loop *Parent,
1379                                                  ValueToValueMapTy &VM,
1380                                                  bool IsSubloop) {
1381   Loop &New = *LI.AllocateLoop();
1382   if (Parent)
1383     Parent->addChildLoop(&New);
1384   else
1385     LI.addTopLevelLoop(&New);
1386   LPMAddNewLoop(&New, IsSubloop);
1387 
1388   // Add all of the blocks in Original to the new loop.
1389   for (auto *BB : Original->blocks())
1390     if (LI.getLoopFor(BB) == Original)
1391       New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), LI);
1392 
1393   // Add all of the subloops to the new loop.
1394   for (Loop *SubLoop : *Original)
1395     createClonedLoopStructure(SubLoop, &New, VM, /* IsSubloop */ true);
1396 
1397   return &New;
1398 }
1399 
1400 bool LoopConstrainer::run() {
1401   BasicBlock *Preheader = nullptr;
1402   LatchTakenCount = SE.getExitCount(&OriginalLoop, MainLoopStructure.Latch);
1403   Preheader = OriginalLoop.getLoopPreheader();
1404   assert(!isa<SCEVCouldNotCompute>(LatchTakenCount) && Preheader != nullptr &&
1405          "preconditions!");
1406 
1407   OriginalPreheader = Preheader;
1408   MainLoopPreheader = Preheader;
1409 
1410   bool IsSignedPredicate = MainLoopStructure.IsSignedPredicate;
1411   Optional<SubRanges> MaybeSR = calculateSubRanges(IsSignedPredicate);
1412   if (!MaybeSR) {
1413     LLVM_DEBUG(dbgs() << "irce: could not compute subranges\n");
1414     return false;
1415   }
1416 
1417   SubRanges SR = *MaybeSR;
1418   bool Increasing = MainLoopStructure.IndVarIncreasing;
1419   IntegerType *IVTy =
1420       cast<IntegerType>(Range.getBegin()->getType());
1421 
1422   SCEVExpander Expander(SE, F.getParent()->getDataLayout(), "irce");
1423   Instruction *InsertPt = OriginalPreheader->getTerminator();
1424 
1425   // It would have been better to make `PreLoop' and `PostLoop'
1426   // `Optional<ClonedLoop>'s, but `ValueToValueMapTy' does not have a copy
1427   // constructor.
1428   ClonedLoop PreLoop, PostLoop;
1429   bool NeedsPreLoop =
1430       Increasing ? SR.LowLimit.has_value() : SR.HighLimit.has_value();
1431   bool NeedsPostLoop =
1432       Increasing ? SR.HighLimit.has_value() : SR.LowLimit.has_value();
1433 
1434   Value *ExitPreLoopAt = nullptr;
1435   Value *ExitMainLoopAt = nullptr;
1436   const SCEVConstant *MinusOneS =
1437       cast<SCEVConstant>(SE.getConstant(IVTy, -1, true /* isSigned */));
1438 
1439   if (NeedsPreLoop) {
1440     const SCEV *ExitPreLoopAtSCEV = nullptr;
1441 
1442     if (Increasing)
1443       ExitPreLoopAtSCEV = *SR.LowLimit;
1444     else if (cannotBeMinInLoop(*SR.HighLimit, &OriginalLoop, SE,
1445                                IsSignedPredicate))
1446       ExitPreLoopAtSCEV = SE.getAddExpr(*SR.HighLimit, MinusOneS);
1447     else {
1448       LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1449                         << "preloop exit limit.  HighLimit = "
1450                         << *(*SR.HighLimit) << "\n");
1451       return false;
1452     }
1453 
1454     if (!Expander.isSafeToExpandAt(ExitPreLoopAtSCEV, InsertPt)) {
1455       LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the"
1456                         << " preloop exit limit " << *ExitPreLoopAtSCEV
1457                         << " at block " << InsertPt->getParent()->getName()
1458                         << "\n");
1459       return false;
1460     }
1461 
1462     ExitPreLoopAt = Expander.expandCodeFor(ExitPreLoopAtSCEV, IVTy, InsertPt);
1463     ExitPreLoopAt->setName("exit.preloop.at");
1464   }
1465 
1466   if (NeedsPostLoop) {
1467     const SCEV *ExitMainLoopAtSCEV = nullptr;
1468 
1469     if (Increasing)
1470       ExitMainLoopAtSCEV = *SR.HighLimit;
1471     else if (cannotBeMinInLoop(*SR.LowLimit, &OriginalLoop, SE,
1472                                IsSignedPredicate))
1473       ExitMainLoopAtSCEV = SE.getAddExpr(*SR.LowLimit, MinusOneS);
1474     else {
1475       LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1476                         << "mainloop exit limit.  LowLimit = "
1477                         << *(*SR.LowLimit) << "\n");
1478       return false;
1479     }
1480 
1481     if (!Expander.isSafeToExpandAt(ExitMainLoopAtSCEV, InsertPt)) {
1482       LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the"
1483                         << " main loop exit limit " << *ExitMainLoopAtSCEV
1484                         << " at block " << InsertPt->getParent()->getName()
1485                         << "\n");
1486       return false;
1487     }
1488 
1489     ExitMainLoopAt = Expander.expandCodeFor(ExitMainLoopAtSCEV, IVTy, InsertPt);
1490     ExitMainLoopAt->setName("exit.mainloop.at");
1491   }
1492 
1493   // We clone these ahead of time so that we don't have to deal with changing
1494   // and temporarily invalid IR as we transform the loops.
1495   if (NeedsPreLoop)
1496     cloneLoop(PreLoop, "preloop");
1497   if (NeedsPostLoop)
1498     cloneLoop(PostLoop, "postloop");
1499 
1500   RewrittenRangeInfo PreLoopRRI;
1501 
1502   if (NeedsPreLoop) {
1503     Preheader->getTerminator()->replaceUsesOfWith(MainLoopStructure.Header,
1504                                                   PreLoop.Structure.Header);
1505 
1506     MainLoopPreheader =
1507         createPreheader(MainLoopStructure, Preheader, "mainloop");
1508     PreLoopRRI = changeIterationSpaceEnd(PreLoop.Structure, Preheader,
1509                                          ExitPreLoopAt, MainLoopPreheader);
1510     rewriteIncomingValuesForPHIs(MainLoopStructure, MainLoopPreheader,
1511                                  PreLoopRRI);
1512   }
1513 
1514   BasicBlock *PostLoopPreheader = nullptr;
1515   RewrittenRangeInfo PostLoopRRI;
1516 
1517   if (NeedsPostLoop) {
1518     PostLoopPreheader =
1519         createPreheader(PostLoop.Structure, Preheader, "postloop");
1520     PostLoopRRI = changeIterationSpaceEnd(MainLoopStructure, MainLoopPreheader,
1521                                           ExitMainLoopAt, PostLoopPreheader);
1522     rewriteIncomingValuesForPHIs(PostLoop.Structure, PostLoopPreheader,
1523                                  PostLoopRRI);
1524   }
1525 
1526   BasicBlock *NewMainLoopPreheader =
1527       MainLoopPreheader != Preheader ? MainLoopPreheader : nullptr;
1528   BasicBlock *NewBlocks[] = {PostLoopPreheader,        PreLoopRRI.PseudoExit,
1529                              PreLoopRRI.ExitSelector,  PostLoopRRI.PseudoExit,
1530                              PostLoopRRI.ExitSelector, NewMainLoopPreheader};
1531 
1532   // Some of the above may be nullptr, filter them out before passing to
1533   // addToParentLoopIfNeeded.
1534   auto NewBlocksEnd =
1535       std::remove(std::begin(NewBlocks), std::end(NewBlocks), nullptr);
1536 
1537   addToParentLoopIfNeeded(makeArrayRef(std::begin(NewBlocks), NewBlocksEnd));
1538 
1539   DT.recalculate(F);
1540 
1541   // We need to first add all the pre and post loop blocks into the loop
1542   // structures (as part of createClonedLoopStructure), and then update the
1543   // LCSSA form and LoopSimplifyForm. This is necessary for correctly updating
1544   // LI when LoopSimplifyForm is generated.
1545   Loop *PreL = nullptr, *PostL = nullptr;
1546   if (!PreLoop.Blocks.empty()) {
1547     PreL = createClonedLoopStructure(&OriginalLoop,
1548                                      OriginalLoop.getParentLoop(), PreLoop.Map,
1549                                      /* IsSubLoop */ false);
1550   }
1551 
1552   if (!PostLoop.Blocks.empty()) {
1553     PostL =
1554         createClonedLoopStructure(&OriginalLoop, OriginalLoop.getParentLoop(),
1555                                   PostLoop.Map, /* IsSubLoop */ false);
1556   }
1557 
1558   // This function canonicalizes the loop into Loop-Simplify and LCSSA forms.
1559   auto CanonicalizeLoop = [&] (Loop *L, bool IsOriginalLoop) {
1560     formLCSSARecursively(*L, DT, &LI, &SE);
1561     simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr, true);
1562     // Pre/post loops are slow paths, we do not need to perform any loop
1563     // optimizations on them.
1564     if (!IsOriginalLoop)
1565       DisableAllLoopOptsOnLoop(*L);
1566   };
1567   if (PreL)
1568     CanonicalizeLoop(PreL, false);
1569   if (PostL)
1570     CanonicalizeLoop(PostL, false);
1571   CanonicalizeLoop(&OriginalLoop, true);
1572 
1573   return true;
1574 }
1575 
1576 /// Computes and returns a range of values for the induction variable (IndVar)
1577 /// in which the range check can be safely elided.  If it cannot compute such a
1578 /// range, returns None.
1579 Optional<InductiveRangeCheck::Range>
1580 InductiveRangeCheck::computeSafeIterationSpace(
1581     ScalarEvolution &SE, const SCEVAddRecExpr *IndVar,
1582     bool IsLatchSigned) const {
1583   // We can deal when types of latch check and range checks don't match in case
1584   // if latch check is more narrow.
1585   auto *IVType = cast<IntegerType>(IndVar->getType());
1586   auto *RCType = cast<IntegerType>(getBegin()->getType());
1587   if (IVType->getBitWidth() > RCType->getBitWidth())
1588     return None;
1589   // IndVar is of the form "A + B * I" (where "I" is the canonical induction
1590   // variable, that may or may not exist as a real llvm::Value in the loop) and
1591   // this inductive range check is a range check on the "C + D * I" ("C" is
1592   // getBegin() and "D" is getStep()).  We rewrite the value being range
1593   // checked to "M + N * IndVar" where "N" = "D * B^(-1)" and "M" = "C - NA".
1594   //
1595   // The actual inequalities we solve are of the form
1596   //
1597   //   0 <= M + 1 * IndVar < L given L >= 0  (i.e. N == 1)
1598   //
1599   // Here L stands for upper limit of the safe iteration space.
1600   // The inequality is satisfied by (0 - M) <= IndVar < (L - M). To avoid
1601   // overflows when calculating (0 - M) and (L - M) we, depending on type of
1602   // IV's iteration space, limit the calculations by borders of the iteration
1603   // space. For example, if IndVar is unsigned, (0 - M) overflows for any M > 0.
1604   // If we figured out that "anything greater than (-M) is safe", we strengthen
1605   // this to "everything greater than 0 is safe", assuming that values between
1606   // -M and 0 just do not exist in unsigned iteration space, and we don't want
1607   // to deal with overflown values.
1608 
1609   if (!IndVar->isAffine())
1610     return None;
1611 
1612   const SCEV *A = NoopOrExtend(IndVar->getStart(), RCType, SE, IsLatchSigned);
1613   const SCEVConstant *B = dyn_cast<SCEVConstant>(
1614       NoopOrExtend(IndVar->getStepRecurrence(SE), RCType, SE, IsLatchSigned));
1615   if (!B)
1616     return None;
1617   assert(!B->isZero() && "Recurrence with zero step?");
1618 
1619   const SCEV *C = getBegin();
1620   const SCEVConstant *D = dyn_cast<SCEVConstant>(getStep());
1621   if (D != B)
1622     return None;
1623 
1624   assert(!D->getValue()->isZero() && "Recurrence with zero step?");
1625   unsigned BitWidth = RCType->getBitWidth();
1626   const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
1627 
1628   // Subtract Y from X so that it does not go through border of the IV
1629   // iteration space. Mathematically, it is equivalent to:
1630   //
1631   //    ClampedSubtract(X, Y) = min(max(X - Y, INT_MIN), INT_MAX).        [1]
1632   //
1633   // In [1], 'X - Y' is a mathematical subtraction (result is not bounded to
1634   // any width of bit grid). But after we take min/max, the result is
1635   // guaranteed to be within [INT_MIN, INT_MAX].
1636   //
1637   // In [1], INT_MAX and INT_MIN are respectively signed and unsigned max/min
1638   // values, depending on type of latch condition that defines IV iteration
1639   // space.
1640   auto ClampedSubtract = [&](const SCEV *X, const SCEV *Y) {
1641     // FIXME: The current implementation assumes that X is in [0, SINT_MAX].
1642     // This is required to ensure that SINT_MAX - X does not overflow signed and
1643     // that X - Y does not overflow unsigned if Y is negative. Can we lift this
1644     // restriction and make it work for negative X either?
1645     if (IsLatchSigned) {
1646       // X is a number from signed range, Y is interpreted as signed.
1647       // Even if Y is SINT_MAX, (X - Y) does not reach SINT_MIN. So the only
1648       // thing we should care about is that we didn't cross SINT_MAX.
1649       // So, if Y is positive, we subtract Y safely.
1650       //   Rule 1: Y > 0 ---> Y.
1651       // If 0 <= -Y <= (SINT_MAX - X), we subtract Y safely.
1652       //   Rule 2: Y >=s (X - SINT_MAX) ---> Y.
1653       // If 0 <= (SINT_MAX - X) < -Y, we can only subtract (X - SINT_MAX).
1654       //   Rule 3: Y <s (X - SINT_MAX) ---> (X - SINT_MAX).
1655       // It gives us smax(Y, X - SINT_MAX) to subtract in all cases.
1656       const SCEV *XMinusSIntMax = SE.getMinusSCEV(X, SIntMax);
1657       return SE.getMinusSCEV(X, SE.getSMaxExpr(Y, XMinusSIntMax),
1658                              SCEV::FlagNSW);
1659     } else
1660       // X is a number from unsigned range, Y is interpreted as signed.
1661       // Even if Y is SINT_MIN, (X - Y) does not reach UINT_MAX. So the only
1662       // thing we should care about is that we didn't cross zero.
1663       // So, if Y is negative, we subtract Y safely.
1664       //   Rule 1: Y <s 0 ---> Y.
1665       // If 0 <= Y <= X, we subtract Y safely.
1666       //   Rule 2: Y <=s X ---> Y.
1667       // If 0 <= X < Y, we should stop at 0 and can only subtract X.
1668       //   Rule 3: Y >s X ---> X.
1669       // It gives us smin(X, Y) to subtract in all cases.
1670       return SE.getMinusSCEV(X, SE.getSMinExpr(X, Y), SCEV::FlagNUW);
1671   };
1672   const SCEV *M = SE.getMinusSCEV(C, A);
1673   const SCEV *Zero = SE.getZero(M->getType());
1674 
1675   // This function returns SCEV equal to 1 if X is non-negative 0 otherwise.
1676   auto SCEVCheckNonNegative = [&](const SCEV *X) {
1677     const Loop *L = IndVar->getLoop();
1678     const SCEV *One = SE.getOne(X->getType());
1679     // Can we trivially prove that X is a non-negative or negative value?
1680     if (isKnownNonNegativeInLoop(X, L, SE))
1681       return One;
1682     else if (isKnownNegativeInLoop(X, L, SE))
1683       return Zero;
1684     // If not, we will have to figure it out during the execution.
1685     // Function smax(smin(X, 0), -1) + 1 equals to 1 if X >= 0 and 0 if X < 0.
1686     const SCEV *NegOne = SE.getNegativeSCEV(One);
1687     return SE.getAddExpr(SE.getSMaxExpr(SE.getSMinExpr(X, Zero), NegOne), One);
1688   };
1689   // FIXME: Current implementation of ClampedSubtract implicitly assumes that
1690   // X is non-negative (in sense of a signed value). We need to re-implement
1691   // this function in a way that it will correctly handle negative X as well.
1692   // We use it twice: for X = 0 everything is fine, but for X = getEnd() we can
1693   // end up with a negative X and produce wrong results. So currently we ensure
1694   // that if getEnd() is negative then both ends of the safe range are zero.
1695   // Note that this may pessimize elimination of unsigned range checks against
1696   // negative values.
1697   const SCEV *REnd = getEnd();
1698   const SCEV *EndIsNonNegative = SCEVCheckNonNegative(REnd);
1699 
1700   const SCEV *Begin = SE.getMulExpr(ClampedSubtract(Zero, M), EndIsNonNegative);
1701   const SCEV *End = SE.getMulExpr(ClampedSubtract(REnd, M), EndIsNonNegative);
1702   return InductiveRangeCheck::Range(Begin, End);
1703 }
1704 
1705 static Optional<InductiveRangeCheck::Range>
1706 IntersectSignedRange(ScalarEvolution &SE,
1707                      const Optional<InductiveRangeCheck::Range> &R1,
1708                      const InductiveRangeCheck::Range &R2) {
1709   if (R2.isEmpty(SE, /* IsSigned */ true))
1710     return None;
1711   if (!R1)
1712     return R2;
1713   auto &R1Value = R1.value();
1714   // We never return empty ranges from this function, and R1 is supposed to be
1715   // a result of intersection. Thus, R1 is never empty.
1716   assert(!R1Value.isEmpty(SE, /* IsSigned */ true) &&
1717          "We should never have empty R1!");
1718 
1719   // TODO: we could widen the smaller range and have this work; but for now we
1720   // bail out to keep things simple.
1721   if (R1Value.getType() != R2.getType())
1722     return None;
1723 
1724   const SCEV *NewBegin = SE.getSMaxExpr(R1Value.getBegin(), R2.getBegin());
1725   const SCEV *NewEnd = SE.getSMinExpr(R1Value.getEnd(), R2.getEnd());
1726 
1727   // If the resulting range is empty, just return None.
1728   auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
1729   if (Ret.isEmpty(SE, /* IsSigned */ true))
1730     return None;
1731   return Ret;
1732 }
1733 
1734 static Optional<InductiveRangeCheck::Range>
1735 IntersectUnsignedRange(ScalarEvolution &SE,
1736                        const Optional<InductiveRangeCheck::Range> &R1,
1737                        const InductiveRangeCheck::Range &R2) {
1738   if (R2.isEmpty(SE, /* IsSigned */ false))
1739     return None;
1740   if (!R1)
1741     return R2;
1742   auto &R1Value = R1.value();
1743   // We never return empty ranges from this function, and R1 is supposed to be
1744   // a result of intersection. Thus, R1 is never empty.
1745   assert(!R1Value.isEmpty(SE, /* IsSigned */ false) &&
1746          "We should never have empty R1!");
1747 
1748   // TODO: we could widen the smaller range and have this work; but for now we
1749   // bail out to keep things simple.
1750   if (R1Value.getType() != R2.getType())
1751     return None;
1752 
1753   const SCEV *NewBegin = SE.getUMaxExpr(R1Value.getBegin(), R2.getBegin());
1754   const SCEV *NewEnd = SE.getUMinExpr(R1Value.getEnd(), R2.getEnd());
1755 
1756   // If the resulting range is empty, just return None.
1757   auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
1758   if (Ret.isEmpty(SE, /* IsSigned */ false))
1759     return None;
1760   return Ret;
1761 }
1762 
1763 PreservedAnalyses IRCEPass::run(Function &F, FunctionAnalysisManager &AM) {
1764   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1765   LoopInfo &LI = AM.getResult<LoopAnalysis>(F);
1766   // There are no loops in the function. Return before computing other expensive
1767   // analyses.
1768   if (LI.empty())
1769     return PreservedAnalyses::all();
1770   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1771   auto &BPI = AM.getResult<BranchProbabilityAnalysis>(F);
1772 
1773   // Get BFI analysis result on demand. Please note that modification of
1774   // CFG invalidates this analysis and we should handle it.
1775   auto getBFI = [&F, &AM ]()->BlockFrequencyInfo & {
1776     return AM.getResult<BlockFrequencyAnalysis>(F);
1777   };
1778   InductiveRangeCheckElimination IRCE(SE, &BPI, DT, LI, { getBFI });
1779 
1780   bool Changed = false;
1781   {
1782     bool CFGChanged = false;
1783     for (const auto &L : LI) {
1784       CFGChanged |= simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr,
1785                                  /*PreserveLCSSA=*/false);
1786       Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1787     }
1788     Changed |= CFGChanged;
1789 
1790     if (CFGChanged && !SkipProfitabilityChecks) {
1791       PreservedAnalyses PA = PreservedAnalyses::all();
1792       PA.abandon<BlockFrequencyAnalysis>();
1793       AM.invalidate(F, PA);
1794     }
1795   }
1796 
1797   SmallPriorityWorklist<Loop *, 4> Worklist;
1798   appendLoopsToWorklist(LI, Worklist);
1799   auto LPMAddNewLoop = [&Worklist](Loop *NL, bool IsSubloop) {
1800     if (!IsSubloop)
1801       appendLoopsToWorklist(*NL, Worklist);
1802   };
1803 
1804   while (!Worklist.empty()) {
1805     Loop *L = Worklist.pop_back_val();
1806     if (IRCE.run(L, LPMAddNewLoop)) {
1807       Changed = true;
1808       if (!SkipProfitabilityChecks) {
1809         PreservedAnalyses PA = PreservedAnalyses::all();
1810         PA.abandon<BlockFrequencyAnalysis>();
1811         AM.invalidate(F, PA);
1812       }
1813     }
1814   }
1815 
1816   if (!Changed)
1817     return PreservedAnalyses::all();
1818   return getLoopPassPreservedAnalyses();
1819 }
1820 
1821 bool IRCELegacyPass::runOnFunction(Function &F) {
1822   if (skipFunction(F))
1823     return false;
1824 
1825   ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1826   BranchProbabilityInfo &BPI =
1827       getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
1828   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1829   auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1830   InductiveRangeCheckElimination IRCE(SE, &BPI, DT, LI);
1831 
1832   bool Changed = false;
1833 
1834   for (const auto &L : LI) {
1835     Changed |= simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr,
1836                             /*PreserveLCSSA=*/false);
1837     Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1838   }
1839 
1840   SmallPriorityWorklist<Loop *, 4> Worklist;
1841   appendLoopsToWorklist(LI, Worklist);
1842   auto LPMAddNewLoop = [&](Loop *NL, bool IsSubloop) {
1843     if (!IsSubloop)
1844       appendLoopsToWorklist(*NL, Worklist);
1845   };
1846 
1847   while (!Worklist.empty()) {
1848     Loop *L = Worklist.pop_back_val();
1849     Changed |= IRCE.run(L, LPMAddNewLoop);
1850   }
1851   return Changed;
1852 }
1853 
1854 bool
1855 InductiveRangeCheckElimination::isProfitableToTransform(const Loop &L,
1856                                                         LoopStructure &LS) {
1857   if (SkipProfitabilityChecks)
1858     return true;
1859   if (GetBFI) {
1860     BlockFrequencyInfo &BFI = (*GetBFI)();
1861     uint64_t hFreq = BFI.getBlockFreq(LS.Header).getFrequency();
1862     uint64_t phFreq = BFI.getBlockFreq(L.getLoopPreheader()).getFrequency();
1863     if (phFreq != 0 && hFreq != 0 && (hFreq / phFreq < MinRuntimeIterations)) {
1864       LLVM_DEBUG(dbgs() << "irce: could not prove profitability: "
1865                         << "the estimated number of iterations basing on "
1866                            "frequency info is " << (hFreq / phFreq) << "\n";);
1867       return false;
1868     }
1869     return true;
1870   }
1871 
1872   if (!BPI)
1873     return true;
1874   BranchProbability ExitProbability =
1875       BPI->getEdgeProbability(LS.Latch, LS.LatchBrExitIdx);
1876   if (ExitProbability > BranchProbability(1, MinRuntimeIterations)) {
1877     LLVM_DEBUG(dbgs() << "irce: could not prove profitability: "
1878                       << "the exit probability is too big " << ExitProbability
1879                       << "\n";);
1880     return false;
1881   }
1882   return true;
1883 }
1884 
1885 bool InductiveRangeCheckElimination::run(
1886     Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop) {
1887   if (L->getBlocks().size() >= LoopSizeCutoff) {
1888     LLVM_DEBUG(dbgs() << "irce: giving up constraining loop, too large\n");
1889     return false;
1890   }
1891 
1892   BasicBlock *Preheader = L->getLoopPreheader();
1893   if (!Preheader) {
1894     LLVM_DEBUG(dbgs() << "irce: loop has no preheader, leaving\n");
1895     return false;
1896   }
1897 
1898   LLVMContext &Context = Preheader->getContext();
1899   SmallVector<InductiveRangeCheck, 16> RangeChecks;
1900 
1901   for (auto BBI : L->getBlocks())
1902     if (BranchInst *TBI = dyn_cast<BranchInst>(BBI->getTerminator()))
1903       InductiveRangeCheck::extractRangeChecksFromBranch(TBI, L, SE, BPI,
1904                                                         RangeChecks);
1905 
1906   if (RangeChecks.empty())
1907     return false;
1908 
1909   auto PrintRecognizedRangeChecks = [&](raw_ostream &OS) {
1910     OS << "irce: looking at loop "; L->print(OS);
1911     OS << "irce: loop has " << RangeChecks.size()
1912        << " inductive range checks: \n";
1913     for (InductiveRangeCheck &IRC : RangeChecks)
1914       IRC.print(OS);
1915   };
1916 
1917   LLVM_DEBUG(PrintRecognizedRangeChecks(dbgs()));
1918 
1919   if (PrintRangeChecks)
1920     PrintRecognizedRangeChecks(errs());
1921 
1922   const char *FailureReason = nullptr;
1923   Optional<LoopStructure> MaybeLoopStructure =
1924       LoopStructure::parseLoopStructure(SE, *L, FailureReason);
1925   if (!MaybeLoopStructure) {
1926     LLVM_DEBUG(dbgs() << "irce: could not parse loop structure: "
1927                       << FailureReason << "\n";);
1928     return false;
1929   }
1930   LoopStructure LS = *MaybeLoopStructure;
1931   if (!isProfitableToTransform(*L, LS))
1932     return false;
1933   const SCEVAddRecExpr *IndVar =
1934       cast<SCEVAddRecExpr>(SE.getMinusSCEV(SE.getSCEV(LS.IndVarBase), SE.getSCEV(LS.IndVarStep)));
1935 
1936   Optional<InductiveRangeCheck::Range> SafeIterRange;
1937   Instruction *ExprInsertPt = Preheader->getTerminator();
1938 
1939   SmallVector<InductiveRangeCheck, 4> RangeChecksToEliminate;
1940   // Basing on the type of latch predicate, we interpret the IV iteration range
1941   // as signed or unsigned range. We use different min/max functions (signed or
1942   // unsigned) when intersecting this range with safe iteration ranges implied
1943   // by range checks.
1944   auto IntersectRange =
1945       LS.IsSignedPredicate ? IntersectSignedRange : IntersectUnsignedRange;
1946 
1947   IRBuilder<> B(ExprInsertPt);
1948   for (InductiveRangeCheck &IRC : RangeChecks) {
1949     auto Result = IRC.computeSafeIterationSpace(SE, IndVar,
1950                                                 LS.IsSignedPredicate);
1951     if (Result) {
1952       auto MaybeSafeIterRange =
1953           IntersectRange(SE, SafeIterRange, Result.value());
1954       if (MaybeSafeIterRange) {
1955         assert(!MaybeSafeIterRange.value().isEmpty(SE, LS.IsSignedPredicate) &&
1956                "We should never return empty ranges!");
1957         RangeChecksToEliminate.push_back(IRC);
1958         SafeIterRange = MaybeSafeIterRange.value();
1959       }
1960     }
1961   }
1962 
1963   if (!SafeIterRange)
1964     return false;
1965 
1966   LoopConstrainer LC(*L, LI, LPMAddNewLoop, LS, SE, DT, SafeIterRange.value());
1967   bool Changed = LC.run();
1968 
1969   if (Changed) {
1970     auto PrintConstrainedLoopInfo = [L]() {
1971       dbgs() << "irce: in function ";
1972       dbgs() << L->getHeader()->getParent()->getName() << ": ";
1973       dbgs() << "constrained ";
1974       L->print(dbgs());
1975     };
1976 
1977     LLVM_DEBUG(PrintConstrainedLoopInfo());
1978 
1979     if (PrintChangedLoops)
1980       PrintConstrainedLoopInfo();
1981 
1982     // Optimize away the now-redundant range checks.
1983 
1984     for (InductiveRangeCheck &IRC : RangeChecksToEliminate) {
1985       ConstantInt *FoldedRangeCheck = IRC.getPassingDirection()
1986                                           ? ConstantInt::getTrue(Context)
1987                                           : ConstantInt::getFalse(Context);
1988       IRC.getCheckUse()->set(FoldedRangeCheck);
1989     }
1990   }
1991 
1992   return Changed;
1993 }
1994 
1995 Pass *llvm::createInductiveRangeCheckEliminationPass() {
1996   return new IRCELegacyPass();
1997 }
1998